MDF-YOLO: a context-modulated deformable feature network for accurate small-object detection in UAV images
TL;DR
Experiments show that MDF-YOLO provides a favorable balance between detection accuracy and model complexity and exhibits potential for practical UAV-based object detection applications, and category-wise evaluation further demonstrates consistent AP@50 improvements across all 10 object categories.
Abstract
Accurate detection and localization of small objects in unmanned aerial vehicle (UAV) images are essential for traffic monitoring, urban management, and emergency response. However, UAV imagery usually contains dense object distributions, complex backgrounds, illumination variations, and substantial scale changes, making small objects difficult to distinguish and localize accurately. To address these challenges, we propose MDF-YOLO, a context-modulated deformable feature network based on YOLOv12 for UAV small-object detection. First, a context-modulated deformable large-kernel attention module is introduced into the backbone to enhance geometric adaptability and contextual representation by incorporating large-kernel contextual information into the deformable sampling-weight generation process. Second, an rectangular self-calibration module (RCM)-based rectangular feature calibration module is embedded in the neck to strengthen foreground-aware multi-scale feature fusion and suppress background interference. Third, Focaler-intersection over union (IoU) is adopted as the bounding-box regression loss to improve the localization of difficult samples. Experiments on the VisDrone2019-DET dataset show that, compared with YOLOv12n, MDF-YOLO improves precision from 52.7% to 56.0%, recall from 40.9% to 43.8%, and mAP@50 from 43.1% to 46.0%, corresponding to gains of 3.3, 2.9, and 2.9 percentage points, respectively. MDF-YOLO contains 3.47M parameters and requires 8.50 floating point operations, maintaining a relatively compact model structure. Category-wise evaluation further demonstrates consistent AP@50 improvements across all 10 object categories, with particularly noticeable gains for visually weak and frequently small instances. These results indicate that MDF-YOLO provides a favorable balance between detection accuracy and model complexity and exhibits potential for practical UAV-based object detection applications.